Systematically designing better instance counting

models on cell images with Neural Arithmetic

Logic Units
Ashish Rana∗ , Taranveer Singh† , Harpreet Singh‡ , Neeraj Kumar§ and Prashant Singh Rana¶
Department of Computer Science, TIET Patiala, Punjab, India.
Email: ∗ {arana be15, tsingh me17, harpreet.s, neeraj.kumar and prashant.singh }@thapar.edu

Abstract—The big problem for neural network models which computation and performs relatively better than non linear
arXiv:2004.06674v2 [cs.LG] 15 Jun 2020

are trained to count instances is that whenever testing range data activation functions for arithmetic operations. This numerical
goes to higher counts than training range data generalization bias of learning computations makes them excellent choice for
error increases i.e. prediction error for images that are outside
training range increases. Consider the case of automating cell counting tasks which are essentially is an increment addition
counting process where more dense images with higher cell operation only. Deep learning models generally take either
counts are commonly encountered as compared to images used in segmentation approach with explicit counting trainer or end-
training data. By making better predictions for higher ranges of to-end counting via a regression loss. In this paper we will
cell count we are aiming to create better generalization systems go through the latter approach [4] in detail for automation of
for cell counting. With architecture proposal of neural arithmetic
logic units (NALU) for arithmetic operations, task of counting cell counting process. As cell counting is cumbersome task
has become feasible for higher numeric ranges which were not and dense cell images with higher cell counts containing data
included in training data with better accuracy. In our study we outside training numeric range are common in real world sce-
incorporate these units in already existing Fully Convolutional narios. Achieving true cell automation with less generalization
Regression Network (FCRN) and U-Net architectures in the form errors is the prime objective of this paper.
of residual concatenated layers. We carried out a systematic
comparative study with the newly proposed changes and earlier
base architectures. This comparative study results are evaluated
in terms of optimizing regression loss across cell density map
over an image. We achieved better results in cell counting
tasks with our newly proposed architectures having residual
layer concatenation connections. We further validated our results
on custom created high count dataset created from BBBC005
synthetic cell count dataset and obtained even better results
where testing images have higher counts. These results confirm
that above mentioned numerically biased units does help models
to learn numeric quantities for better generalization results on
high count data.
Index Terms—Neural Arithmetic Logic Units, Cell Counting,
Fully Convolutional Regression Networks, U-net.

I. I NTRODUCTION
Ability to generalize concepts is fundamental component
of intelligence and core for designing smart systems [1],
[2]. Neural networks simulates this behavior with hierarchical
learning of concepts. When it comes to automation, counting is
an important task from machine vision application [3] to cell
counting [4]. While neural networks manipulates numerical
quantities but it is not associated with systematic generaliza-
tion [5], [6]. These networks fail to generalize as evident from
high generalization error while predicting quantities that lie
outside the training numerical range [7]. This highlights mem-
orization behavior in neural networks instead of generalization
abilities for a given task. This is especially problematic for cell
counting tasks as images with higher cell counts that were not
part of training are common to encounter in real applications.
Neural accumulators (NAC) and neural arithmetic logic
units (NALU) [7] are biased to learn systematic numerical
Fig. 1. Abstract representation of proposed modified architecture (Right) with
either NAC or NALU as compared to previous Fully Convolution Regression
Networks (Left). Here, as clearly specified with dimensional analysis of tensor
blocks (Right) the layers in the newly proposed model are concatenated and
again squeezed into same size as earlier base model architecture.
In regression loss approach Fully convolutional regression
networks [2] and U-net [8] architectures learns mapping be-
tween an image I(x) and a density map D(x), given by F : I(x)
→ D(x) (I ∈ Rm × n , D ∈ Rm × n ) for a m × n pixel image.
Later on, variations of these base architectures with different
activation functions are implemented in this paper and their
prediction performance is compared with newly proposed
NAC/NALU concatenated architectures of FCRN and U-net.
The concatenated connections adds numerical bias to layers
of the network and behaves in similar manner like ResNets
[9]. But, instead of addition of inputs from previous layer the
input from previous is passed through these numerical units
for capturing numerical bias and concatenated with the main
network layer as shown in figure 1. On surface it appears
that these proposed architectural changes leads to accuracy
improvements due to increased model capacity with these
numerically biased units adding more parameters for learning.
But, our results with custom high count dataset for testing
created from BBBC005 [10] reflects increased generaliza-
tion counting abilities. As for higher cell count images our
model gives higher relative improvement in mean absolute
error(MAE).
Our concatenation based residual architecture utilizes the
fundamentals of batch normalization like specified identity
mapping architecture [11] in ResNets. But, instead of using
convolution operation directly this network leverages numeri-
cal bias information obtained from NAC and NALU operations
applied on input layer and then finally uses convolution
operation on the concatenated layer. Before and after this
concatenation of this numerical bias learning operation, batch
normalization is carried out and output of this operation is
added back again to our next main network layer, as shown
in figure 1.
With means of this paper we introduce changes in current
regression based model architectures for end-to-end counter
training and produce systems with improved accuracy. Also,
we validate our trained models on a different specially tai-
lored validation dataset with approximately seven times higher
counts of cells as compared to training dataset created from
BBBC005 synthetic cell dataset [10]. Overall, experimental
results demonstrates that supplementing the base architectures
with NAC and NALU helped in achieving better results and
improved relative MAE for higher count test images.
II. R ELATED W ORK
Intuitive numerical understanding is important in learning
and by adjunct important in deep learning [1] for creating
better models with higher generalization capabilities. Counting
objects [12]–[16] in given image is a widely studied task.
Trained models for counting tasks either use a deep learning
model to segment instances of given object then count them in
a post-processing step [17] or learn end-to-end predict count
via a regression loss [4]. Networks like Count-ception [18]
added the concept of average over redundant predictions with
its deep Inception family network. Also, recent architectures
like ResNets [19], Highway Networks [20] and Densenets [21]
advocate linear connections like Count-ception to promote
better learning bias. Such models have better performance,
though additional computational overhead due to increased
depth of given architectures do arise. Our work highlights the
generalization capabilities of the network, that extrapolate well
on unseen parts of solution space which highlights underlying
structure of behavior governing-equations [22].
We introduce architectural changes in models that learns and
preserves input information which is numerically biased with
reference to input layers. It is somewhat similar to ResNets [9],
which are easier to optimize and gain accuracy with increasing
depth. With our experiments we aim to highlight that models
with numerically biased concatenated residual functions helps
in achieving better results with their addition in the form of
a comparative study with original architectures. Also, with
our results in this paper we demonstrate with our results that
backpropagation learns this numerical bias without any explicit
numeric quantity being provided as input implying that better
computer vision counters can be trained with this module when
added to existing convolutional neural network architectures.
Density based estimation doesnt require prior object detec-
tion or segmentation [12], [15], [23]. In previous years, several
works have investigated this approach. In [15], the problem
is stated as density estimation with a supervised learning
algorithm, D(x) = cT φ(x), where D(x) represents ground-truth
density map, and φ(x) represents local features and parameters
c are learned by minimizing the error between predicted and
true density with quadratic programming over all possible
sub-windows. In [23], regression forest is used to exploit
patch-based idea for learning structured labels, then for new
input image density map is estimated averaged over structured
patch-based predictions. Also, in [12] an algorithm is used that
allows fast interactive counting with ridge regression.
Fig. 2. The left column represents training image in first row and corre-
sponding annotated image in second row from fluorescent synthetic dataset
[29]. The second column of highlights corresponding zoomed view for better
visual comprehension of the images in the dataset.
Cell counting [15] problem is classified into supervised
learning problem that learns mapping between an image I(x)
and a density map D(x), denoted by F : I(x) → D(x) (I ∈
Rm × n , D ∈ Rm × n ) for a m × n pixel image, see figure
2. Density function D(x) function is defined on pixels in
given image, integrating this map over an image region gives
an estimate of number of cells in that region. CNNs [24],
[25] are quite popular in the bio-medical imaging because
of their simple architecture and achieve great results. Like
in mitosis detection [26], neuronal membrane segmentation
[27] and analysis of C. elegans embryos development [28].
Previously, fully convolutional regression networks (FCRNs)
and Count-ception have given state-of-the-art results in cell
counting, with potential for cell detection of overlapping cells.
Also U-Nets [8], typically fully convolutional network uses
a modified version of architecture proposed by Ciresan et
al. [27] as latter is slow and trade-off between localization
and use of context are present. In U-Nets pooling operations
are replaced by upsampling operations to supplement usual
contracting network. For localization high resolution features Fig. 3. Left: NACs are biased towards learning [-1, 0, 1] as highlighted
with large plateau regions around these values. This means its outputs are
from contracting path are combined with unsampled output. either addition or subtraction of input vectors not scaling. Right: Approximate
Based on this information a successive convolution layer then surface curve of NALU with some dimensional constraints for 3-D plotting.
learn to assemble more precise output. For our experimentation It highlights the ability to scale, along with earlier numerical biases around
[-1, 0, 1] as shown with plateau region surfaces.
we selected FCRN and U-net based on simplicity and relative
similarity in their architectures with the difference being that
U-net uses inputs from previous layer for better localization. B. Comparative analysis of addition operation
Here, we use neural networks with NACs/NALUs and mul-
III. E XPERIMENTS tilayer perceptrons (MLP) with different activation functions
but same structures. These are trained with two randomly
In this section first we conceptually explore NACs and
generated inputs from uniform distribution a and b with each
NALUs. Then compare their addition capabilities with multi-
having 214 data points for training. Prediction capabilities on
layer perceptrons equipped with different activation functions.
test data with values ranging up to 10 times the training
With this study we aim to select NAC/NALU designed vari-
range are evaluated as part of this experiment. Refer figure
ants which can best approximate the counting behavior and
4 to observe architecture for both these trained models in this
compare them with standard FCRN and U-net neural network
comparative study.
architecture for regression loss approach in the following
experiment done on synthetic dataset [29]. For validation
of counting generalization capabilities, our trained models
are tested against different synthetic cell image dataset with
approximately seven times higher counts than training data.

A. Visual understanding of NACs and NALUs

NACs [7] supports accumulation of numerical quantities ad-
ditively, a desirable bias for linear exploration while counting.
It is special type of linear layer with transformation matrix
W being continuous and differentiable parameterization for
gradient descent. W = tanh(Ŵ ) σ(M̂ ) consists of elements
in [-1, 1] with bias close to 1, 0, and 1. See figure 3 for ideation
of this concept with following equations for NAC: a = Wx,
W = tanh(Ŵ ) σ(M̂ ) where Ŵ , M̂ are learning parameters
and W is transformation matrix.
For complex mathematical operations like multiplication
and division we use NALUs. It uses weighted sum of two
sub-cells, one for addition or subtraction and another of
multiply, division or power functions. It demonstrates that
neural accumulators (NACs) can be extended for learning
scaling operations with gate-controlled sub-operations. See
figure 3 for ideation of this concept with following equations
for NALU: y = g a + (1-g) m; m = expW(log( |x| + )), g
= σ(Gx) where m is subcell that operates in log space and g
is learned gate, both contains learning parameters.
Fig. 4. The model on left represents neural network with NAC or NALU units
and on right, a MLP which is trained with different activation functions. Each
variant of above stated models have two inputs, three hidden units, one output.
Comparative analysis is summarized in table I with MAE as
an accuracy measure for MLP variants, NAC, NALU and its
variants with changed learned gate for extrapolation. Also, in
NALU-Tanh and NALU-Hard Sigmoid the learning gate g’s is
changed to observe any improvements in NALU’s performance
based on this change.
From these results stated in table I and visualized in figure
5 we conclude that Linear, LeakyReLU, ReLU activations and
NAC, NALU, NALU-Tanh modules were the top performers in
extrapolation task for numeric addition operation task. Hence,
these top performers are used further in cell-counting task on
synthetic dataset for learning end-to-end counting mechanism.
 TABLE I is comprised of 600 images have a corresponding foreground
C OMPARATIVE RESULT SUMMARIZATION FOR MULTIPLE MODELS mask which alters the focus on these images and ground truth
Layer Configuration/Activations Mean Absolute Error (a+b) images are completely in-focused version before any guassian
Linear MLP 3.63 × 10−6 filter application. We take a subset of this dataset with highly
Sigmoid MLP 29.830 focused F1 images only and their corresponding ground truth
Tanh MLP 15.743 image for validating our model. And after that we coalesce
ELu MLP 0.019 each image 16 times in 4x4 grid manner with random vertical
ReLU MLP 0.001 and horizontal flips to create a final high cell count images for
Leaky ReLU MLP 9.83 × 10−4 our validation dataset, see figure 6. The ground truth images
PReLU MLP 0.001 are accordingly changed with same randomness and they are
NAC 2.70 × 10−6 eroded also to match the fluorescent synthetic dataset on which
NALU 2.71 × 10−6 the models are trained on.
NALU-Hard Sigmoid 3.24 × 10−6
NALU-Tanh 3.18 × 10−6

Fig. 6. Left: Original image from BBBC005 dataset. Right: Repeated image
generated from a 4X4 grid repetition operation with random horizontal and
vertical flips of base image shown in left.

Data augmentation with elastic deformations to training im-

ages is applied for teaching network the desired invariance and
Fig. 5. Above visualization of MAE and log10 (MAE) for different models robustness properties, like specified in figure 7. These elastic
learning identity mapping, demonstrates that our standard activation functions
like sigmoid and tanh doesn’t perform that well with higher data ranges during
deformations are introduced in the form of angular shear in the
testing. Whereas, NACs and NALUs clearly overpowers the identity learning training images. Translation and rotation invariance along with
task and linear, relu based activation functions somewhat manages to provide robustness to gray value variations and deformations is main
acceptable results.
focus of augmentation process for microscopic images. Disfig-
urement using random displacement vectors on a coarse 3x3
C. Cell counting experiment grid are also generated. These data augmentation techniques
1 especially are helpful for our custom data which is created
just by repeating the original image in order to supplement a
In this experiment section the first subsections elaborates
more robust dataset for the model to train on.
the datasets used for training and validation of our trained
models, plus the data augmentation techniques used in our
experiment. After that we elaborate onto different architectures
used for training having different activation layers on standard
architectures and residual concatenated connection modules on
modified proposed model architectures.
1) Datasets and data augmentation: Synthetic dataset
which is generated by system [29]. 200 highly-realistic syn-
thetic fluorescence microscopic images of bacterial cells are
used for experimentation with a 75/25 train-test split for
training each model architecture and its variants. Images are
having average of 17464 cells.
For validation of trained models and checking true gener-
alization capabilities we use BBBC005 from the the Broad Fig. 7. The above figure illustrates different cell augmentation techniques in
Institutes Bioimage Benchmark Collection [10]. This dataset action like elastic deformations, random rotations and plane shifting for the
input images adding more robustness to the trained model.
1 Code repository: https://github.com/ashishrana160796/nalu-cell-counting
 2) Defining regression task and architecture details: In
training dataset ground truth is provided as dot annotation
corresponding to each cell image. For training, dot annotations
are represented by Gaussian and density surface D(x) which is
formed from superposition of Gaussians. The optimization task
is to regress density surface from corresponding image I(x).
This is achieved by training convolutional neural networks
(CNN) using mean square error between output heat map and
target density surface as the loss function. Hence, at inference
given an input I(x), the model predicts density heat map D(x).
Fig. 8. Network Architecture Left: FCRN with 3x3 convolution operations.
Network Architecture Right: FCRN added with residual concatenation con-
nections of Ni numerically biased units. Also, for dimensional compatibility
Ni residual layers are fed to corresponding main network layers and after that
normalized for regularization. At last, 1x1 Conv operation output in the form
of density map of result image is generated .
Conv + BN + Pool: A 3x3 convolutional operation with batch normalization
regularization and 2x2 max pooling layer.
Unsample+Conv + BN: Unsampling the image and then applying convolu-
tional operation with batch normalization regularization.
Concat: Feature maps concatenated along depth dimension.
Ni: NAC or Variants of NALUs as residual concatenated connection.
FCRNs are inspired from VGG-net, we only used small
kernels of size 3x3 pixels for designing our network. Feature
maps are increased for avoiding spatial information loss.
Activation layers like convolution-ReLU-Pooling are popular
in CNN architectures [24]. Here, we have altered these layers
to create different models which contains some numerical bias
in the form of residual connections and regularized by batch
normalization. The first layers contains convolutions-pooling
operations, then we undo spatial reduction by upsampling op-
erations for end-to-end model training. Also, for dimensional
compatibility of residual NAC or NALU modules we append
their output to corresponding main network layer across depth
dimension and batch normalize the output after appropriate
convolutional operation. See figure 8 for comparison between
earlier original model and newly proposed architecture along
with architecture parameter details.
U-net is modification upon the previously discussed FCRN
architecture by having large number of feature channels for
upsampling to propagate context information to high resolu-
tion layers. That makes expansive path almost symmetric to
contracting path yielding a u-shape. Similar to above FCRN’s
optimization problem formulation the evaluation parameters
remains the same. Residual concatenated connection addition
with NACs and NALU units along with batch normalization
is done for adding numerical bias information to the main
network. Also, U-net architecture used in this paper is more
computationally expensive than FCRN having approximately
thrice the number of parameters leading to more feature learn-
ing capacity and is expected to perform better for prediction
tasks. See figure 9 for comparison between earlier original
U-net model and newly proposed architecture along with
parameter details.
For the concatenation of residual connections of these units
the dimensional consistency is maintained by added them
along the depth dimension with the main network. After that
upsampling, convolutional and batch normalization regulariza-
tion operations are applied accordingly for these units to merge
with the base network. FCRNs implementation resembles that
of MatConvNet [30] as upsampling in Keras is implemented
by repeating elements, instead of bilinear sampling. In U-nets,
low-level feature representations are fused during upsampling,
aiming to compensate the information loss due to max pooling.
IV. R ESULTS
MAE is the metric used in this paper for measuring results
for cell counting on the synthetic cell dataset [29] and cus-
tom BBBC005 synthetic modified high cell count validation
dataset.
• Mean Absolute Error (MAE): Mean Absolute Error
(MAE): The mean absolute error is an average of the
difference between the predicted value and true value.
AE = kei k = kyi − xi k (1)
n
X |ei |
M AE = (2)
i=1
n
• Relative Improvement Percentage Relative Improve-
ment Percentage (RIP): Here, in context of this paper
 it defined as percentage improvement in MAE of a
given model with respect to baseline ReLU models for
FCRN and U-net architectures. In below equation, Mr is
MAE from baseline ReLU model and Mi is model under
consideration.

RIP% = ((Mr − Mi )/Mr ) ∗ 100 (3)

Result table II compares baseline FCRN, U-net architectures

with new numerically biased ResNet like connection modules
with NACs and NALUs units under current training setup.
With our setup we able to obtain similar results as mentioned
in earlier reference papers and also we have equipped earlier
model architectures with different regularization activations
as specified in the table. From earlier ReLU implementation
clearly Linear and LeakyReLU activation regularization based
baseline models have performed well. Also, for both model
structures NAC and NALUs residual modules have outper-
formed all the earlier specified regular FCRN architecture. And
similar results are extended by U-net model evaluation where
NALU layer concatenation based U-net outperforms all the
models trained for our experiment.

TABLE II
R ESULT SUMMARIZATION FOR TRAINED MODELS

FCRN-Models MAE U-Net-Models MAE

ReLU 3.43 ReLU 1.78
LeakyReLU 3.39 LeakyReLU 1.74
Linear 3.34 Linear 1.73
NALU-tanh 3.21 NALU/tanh 1.56
NALU 3.17 NALU 1.42
NAC 3.23 NAC 1.63

Result table III compares performance of above trained

models on a new validation dataset containing much higher
cell counts for measuring performance on extrapolation capa-
bilities counting tasks. For validation set we have used 300
images of size 256x256 pixels with cell counts averaging
around 120012. Here also, NAC and NALU based resid-
ual concatenation module based models outperforms earlier
architectures for counting tasks. But, this time the relative
improvement is even more for NALU based FCRN and U-net
models showcasing better generalization abilities of trained
models.

TABLE III
VALIDATING TRAINED MODELS FOR EXTRAPOLATION CELL COUNTING
TASKS
Fig. 9. Network Architecture Top: U-net with 3x3 convolution operations
and increasing depth of filters from 64 to 256 for feature abstraction & FCRN-Models MAE U-Net-Models MAE
learning and after that these layers are fed to upsampling layers. Network ReLU 3.04 ReLU 2.87
Architecture Bottom: U-nets main network added with residual concatenation
LeakyReLU 2.99 LeakyReLU 2.62
connections of Ni numerically biased units and after that batch normalized
for regularization. Similar, operation abbreviations used as stated in figure 8 Linear 2.85 Linear 2.47
NALU-tanh 2.32 NALU-tanh 1.95
NALU 2.27 NALU 1.87
NAC 2.40 NAC 1.92
 Relative improvement in predictions is visualized in figure
10 against ReLU based activation base result for comparison
with other activation layer changes in FCRNs/U-nets and
concatenation layer NALU/NAC residual connection addition
in FCRNs & U-nets. It includes averaged out comparison
from multiple executions of training and testing runs for
both interpolation testing and extrapolation validation counting
tasks for FCRN and U-net variant models with respect to
ReLU based FCRN and U-net model. From, this figure it
is clearly highlighted that models with NAC and NALUs
residual modules have better generalization capabilities for
extrapolation counting tasks i.e. they are better generalizers for
this given cell counting task with increase in relative improve-
ment in prediction as compared to base ReLU implementation.
Here, for measuring generalization capabilities of model in
extrapolation task relative improvement metric is selected as
it bring more perspective to the improved performance on
validation dataset. The performance has shown a general incre-
ment for every model as validation dataset consist of highly
focused images. Hence, measuring relative improvement in
results is more appropriate decision to get understanding of
true improvment in model prediction capabilities.
Fig. 10. Moving towards right in above figure and measuring RIP metric there
is sharp increase with NALU & NAC units based models demonstrating even
more increase in relative improvement with respect to baseline ReLU models
for extrapolation tasks. For interpolation task there is highest 9% improvement
and for extrapolation there is 23% improvement with NAC units as residual
connection.
This figure 10 shows more increase in relative improvement
as we move right towards horizontal axis for both testing
and validation task with extrapolation where in validation
extrapolation task NAC/NALU models from which we can
conclude that trained models are having better generalization
abilities with some learned numerical bias in their trained
weights with which even better predictions for higher count
cells is made.
V. S UMMARY
We were able to demonstrate that addition of newly pro-
posed NAC and NALU units in existing architectures in
the form of residual concatenation connection layer mod-
ules achieves better results. With numerically biased residual
connections, higher accuracy for more dense images having
higher counts of cells is achieved. Hence, producing more
generalized cell counters that provides better predictions for
real life use-cases. Finally, for code implementation details and
other supplementary experimental results refer to this paper’s
github repository.
R EFERENCES
[1] Dehaene, Stanislas. The number sense: How the mind creates mathe-
matics. OUP USA, 2011.
[2] Gallistel, C. Randy. ”Finding numbers in the brain.” Philosophical
Transactions of the Royal Society B: Biological Sciences 373.1740
(2018): 20170119.
[3] Baygin, Mehmet, et al. ”An Image Processing based Object Count-
ing Approach for Machine Vision Application.” arXiv preprint
arXiv:1802.05911 (2018).
[4] Xie, Weidi, J. Alison Noble, and Andrew Zisserman. ”Microscopy cell
counting and detection with fully convolutional regression networks.”
Computer methods in biomechanics and biomedical engineering: Imag-
ing & Visualization 6.3 (2018): 283-292.
[5] Fodor, Jerry A., and Zenon W. Pylyshyn. ”Connectionism and cognitive
architecture: A critical analysis.” Cognition 28.1-2 (1988): 3-71.
[6] Marcus, Gary F. The algebraic mind: Integrating connectionism and
cognitive science. MIT press, 2018.
[7] Trask, Andrew, et al. ”Neural arithmetic logic units.” Advances in Neural
Information Processing Systems. 2018.
[8] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. ”U-net: Con-
volutional networks for biomedical image segmentation.” International
Conference on Medical image computing and computer-assisted inter-
vention. Springer, Cham, 2015.
[9] He, Kaiming, et al. ”Deep residual learning for image recognition.”
Proceedings of the IEEE conference on computer vision and pattern
recognition. 2016.
[10] Ljosa, Vebjorn, Katherine L. Sokolnicki, and Anne E. Carpenter. ”An-
notated high-throughput microscopy image sets for validation.” Nature
methods 9.7 (2012): 637-637.
[11] He, Kaiming, et al. ”Identity mappings in deep residual networks.”
European conference on computer vision. Springer, Cham, 2016.
[12] Arteta, Carlos, et al. ”Interactive object counting.” European conference
on computer vision. Springer, Cham, 2014.
[13] Chan, Antoni B., Zhang-Sheng John Liang, and Nuno Vasconcelos.
”Privacy preserving crowd monitoring: Counting people without people
models or tracking.” 2008 IEEE Conference on Computer Vision and
Pattern Recognition. IEEE, 2008.
[14] Segu, Santi, Oriol Pujol, and Jordi Vitria. ”Learning to count with deep
object features.” Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition Workshops. 2015.
[15] Lempitsky, Victor, and Andrew Zisserman. ”Learning to count objects
in images.” Advances in neural information processing systems. 2010.
[16] Zhang, Cong, et al. ”Cross-scene crowd counting via deep convolutional
neural networks.” Proceedings of the IEEE conference on computer
vision and pattern recognition. 2015.
[17] Hernndez, Carlos X., Mohammad M. Sultan, and Vijay S. Pande. ”Using
Deep Learning for Segmentation and Counting within Microscopy
Data.” arXiv preprint arXiv:1802.10548 (2018).
[18] Paul Cohen, Joseph, et al. ”Count-ception: Counting by fully convo-
lutional redundant counting.” Proceedings of the IEEE International
Conference on Computer Vision. 2017.
[19] He, Kaiming, et al. ”Identity mappings in deep residual networks.”
European conference on computer vision. Springer, Cham, 2016.
[20] Srivastava, Rupesh Kumar, Klaus Greff, and Jrgen Schmidhuber. ”High-
way networks.” arXiv preprint arXiv:1505.00387 (2015).
[21] Huang, Gao, et al. ”Densely connected convolutional networks.” Pro-
ceedings of the IEEE conference on computer vision and pattern
recognition. 2017.
[22] Brunton, Steven L., Joshua L. Proctor, and J. Nathan Kutz. ”Discovering
governing equations from data by sparse identification of nonlinear
dynamical systems.” Proceedings of the National Academy of Sciences
113.15 (2016): 3932-3937.
[23] Fiaschi, Luca, et al. ”Learning to count with regression forest and
structured labels.” Proceedings of the 21st International Conference on
Pattern Recognition (ICPR2012). IEEE, 2012.
[24] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. ”Imagenet
classification with deep convolutional neural networks.” Advances in
neural information processing systems. 2012.
[25] LeCun, Yann, et al. ”Gradient-based learning applied to document
recognition.” Proceedings of the IEEE 86.11 (1998): 2278-2324.
[26] Cirean, Dan C., et al. ”Mitosis detection in breast cancer histology
images with deep neural networks.” International Conference on Medical
Image Computing and Computer-assisted Intervention. Springer, Berlin,
Heidelberg, 2013.
[27] Ciresan, Dan, et al. ”Deep neural networks segment neuronal mem-
branes in electron microscopy images.” Advances in neural information
processing systems. 2012.
[28] Ning, Feng, et al. ”Toward automatic phenotyping of developing em-
bryos from videos.” IEEE Transactions on Image Processing 14 (2005):
1360-1371.
[29] Lehmussola, Antti, et al. ”Computational framework for simulating flu-
orescence microscope images with cell populations.” IEEE transactions
on medical imaging 26.7 (2007): 1010-1016.
[30] Vedaldi, Andrea, and Karel Lenc. ”Matconvnet: Convolutional neural
networks for matlab.” Proceedings of the 23rd ACM international
conference on Multimedia. ACM, 2015.